mine-embed-v1

Training data used to build the experimental embedding model MineAITechnology/mine-embed-v1 (the model; this card is for the dataset of the same name). It has two parts:

1. A tokenized pretraining corpus (GPT-2 BPE) for masked-language-model (MLM) pretraining.
2. A contrastive triplet set (anchor, positive, hard negative) for embedding fine-tuning.

Version 1 (first, experimental test release). This is the first iteration of the data and of the model trained on it. It is published to document an early experiment, not as a finished or production-ready resource. Expect the composition to change in later versions (broader English text, more varied contrastive pairs), and expect models trained only on v1 to score well below established embedding models (see section 5).

Status: research release, mixed licenses. Several sources have their own terms (see section 6). Check each one before any redistribution or commercial use.

1. Contents

Part	Size	Used for
Tokenized corpus	18 train shards (874,248 rows, about 895M tokens) + 1 validation shard (4,393 rows)	MLM pretraining
Contrastive triplets, train	471,400	Contrastive fine-tuning
Contrastive triplets, validation	9,620	Held-out triplet accuracy

Layout (folder and file names as produced by the dataset builder):

tokenized/     *.npy token shards + build_fingerprint.json
contrastive/   pairs_train.jsonl, pairs_val.jsonl + build_fingerprint.json   <!-- TODO: confirm file names -->
tokenizer/     tokenizer_manifest.json

Tokenizer: GPT-2 BPE (gpt2), vocabulary 50,257 plus a <mask> token (id 50257), so the effective vocabulary is 50,258. End-of-text id: 50256 (also used as padding and document separator).

Build fingerprints: tokenized 0db2e603ec2cc6a2 (target 1B tokens, 878,640 windows), contrastive d7706b95d8decdfd.

2. Pretraining corpus mix

Targeted allocation of the tokenized corpus:

Domain	Share	Content
Code and multilingual	45%	GitHub code, English web text, Urdu text, Hinglish bi-text
Math and reasoning	25%	GSM8K, MATH, and generated arithmetic sequences
Security and alignment	20%	Toxicity prompts, prompt-injection and jailbreak samples
Corporate facts	10%	Corporate facts, compliance policies, enterprise retrieval text

Important caveats

• Large parts of the math, security and corporate-fact portions are synthetic or template-generated. The amount of distinct natural text is far below the nominal token count.
• English prose (encyclopedic, scientific, forum, news text) is only a small part of the mix. Models pretrained only on this corpus should not be expected to match general-purpose embedding models on general English benchmarks.
• The security portion contains toxic, abusive and adversarial text by design. Handle accordingly.

3. Contrastive triplets

Each triplet is an anchor, a positive, and a hard negative, for InfoNCE or margin-based losses.

Validation set composition (9,620 triplets):

Source	Count
AllNLI (SNLI + MultiNLI)	8,014
Quora duplicate questions	1,589
Corporate (synthetic/curated)	6
Math/code (synthetic/curated)	9
Security (synthetic/curated)	2

Training set: 471,400 triplets, dominated by AllNLI, with Quora and a small amount of synthetic/curated domain data.

The three synthetic/curated validation domains have too few examples (2 to 9) to measure anything. Do not report accuracy on them.

4. Quality checks

• Exact-match overlap between the train and validation triplet files: 0 (exact string match only).
• Near-duplicate and paraphrase leakage were not checked.
• Validation triplets come from the same sources as training, so triplet accuracy is an in-distribution measure and says little about general embedding quality.

5. Results of a model trained on this data

mine-embed-v1 (57.8M parameters, trained from scratch on this data) scored:

Evaluation	Result
STS-B test (Spearman)	0.686 (0.357 for the MLM-only model before contrastive tuning)
MTEB English subset, 28 tasks, average	37.8
For reference: all-MiniLM-L6-v2 / bge-small-en-v1.5 / GloVe average	61.1 / 64.4 / 45.8

The model is competitive on sentence-similarity tasks such as SICK-R (73.4) and weak on retrieval (4.6), clustering (11.2) and duplicate-question detection. In short: this data teaches fine sentence-pair distinctions (the AllNLI/Quora style) but is too narrow for general topical retrieval. Full tables and per-task scores are in the model repository.

6. Licensing and data terms

This dataset combines material under different licenses. It is not offered under a single open license, and the builder's source metadata is the reference. Verify each source's terms yourself.

Source	Terms to check
SNLI	CC BY-SA 4.0 (share-alike may apply to derivatives)
MultiNLI	Per-genre terms; see the MultiNLI release
Quora Question Pairs	Quora's own terms; commercial use unclear
GitHub code	Per-repository licenses; often copyleft or attribution-required
GSM8K, MATH	See each dataset's license
Toxicity / prompt-injection / jailbreak prompts	See each source's license; contains harmful text
Urdu and Hinglish text
MS MARCO	Research-use terms; only relevant if included (see section 3)
Synthetic and private domain data	Owned by MineAI Technology

Until this review is finished, treat the dataset as research use only.

7. Intended use and limits

• Intended: research on embedding training, reproducing mine-embed-v1, studying contrastive data composition.
• Not suitable as the only data for a general-purpose embedding model (see section 5).
• Not for safety classification or content moderation training without independent review of the security portion.

8. Citation

@misc{mine_embed_v1_2026,
  title        = {mine-embed-v1: pretraining and contrastive data for mine-embed-v1},
  author       = {{MineAI Technology}},
  year         = {2026},
  howpublished = {Hugging Face dataset},
  doi          = {10.57967/hf/10855}
}